fix: ensure permission/stats persist until explicit disconnect (#110)

Also includes cached tokens in cost calculations to provide more accurate billing estimates.

Reviewed-on: #110
Co-authored-by: Naomi Carrigan <commits@nhcarrigan.com>
Co-committed-by: Naomi Carrigan <commits@nhcarrigan.com>

src-tauri/Cargo.lock

@@ -1636,7 +1636,7 @@ checksum = "7f24254aa9a54b5c858eaee2f5bccdb46aaf0e486a595ed5fd8f86ba55232a70"
 
 [[package]]
 name = "hikari-desktop"
-version = "1.2.0"
+version = "1.3.0"
 dependencies = [
  "chrono",
  "dirs 5.0.1",

src-tauri/src/achievements.rs

@@ -2329,6 +2329,10 @@ mod tests {
             context_utilisation_percent: 0.0,
             potential_cache_hits: 0,
             potential_cache_savings_tokens: 0,
+            current_request_input: None,
+            current_request_output_chars: 0,
+            current_request_thinking_chars: 0,
+            current_request_tools: Vec::new(),
             achievements: AchievementProgress::new(),
         }
     }

src-tauri/src/stats.rs

@@ -154,6 +154,16 @@ pub struct UsageStats {
     // Achievement tracking
     #[serde(skip)]
     pub achievements: AchievementProgress,
+
+    // Track current in-flight request for cost estimation on interrupt
+    #[serde(skip)]
+    pub current_request_input: Option<String>,
+    #[serde(skip)]
+    pub current_request_output_chars: u64,
+    #[serde(skip)]
+    pub current_request_thinking_chars: u64,
+    #[serde(skip)]
+    pub current_request_tools: Vec<String>,
 }
 
 impl UsageStats {
@@ -163,13 +173,26 @@ impl UsageStats {
         stats
     }
 
-    pub fn add_usage(&mut self, input_tokens: u64, output_tokens: u64, model: &str) {
+    pub fn add_usage(
+        &mut self,
+        input_tokens: u64,
+        output_tokens: u64,
+        model: &str,
+        cache_creation_tokens: Option<u64>,
+        cache_read_tokens: Option<u64>,
+    ) {
         self.total_input_tokens += input_tokens;
         self.total_output_tokens += output_tokens;
         self.session_input_tokens += input_tokens;
         self.session_output_tokens += output_tokens;
 
-        let cost = calculate_cost(input_tokens, output_tokens, model);
+        let cost = calculate_cost(
+            input_tokens,
+            output_tokens,
+            model,
+            cache_creation_tokens,
+            cache_read_tokens,
+        );
         self.total_cost_usd += cost;
         self.session_cost_usd += cost;
 
@@ -439,6 +462,10 @@ impl UsageStats {
             potential_cache_hits: self.potential_cache_hits,
             potential_cache_savings_tokens: self.potential_cache_savings_tokens,
             achievements: AchievementProgress::new(), // Dummy for copy
+            current_request_input: None, // Don't copy tracking fields
+            current_request_output_chars: 0,
+            current_request_thinking_chars: 0,
+            current_request_tools: Vec::new(),
         };
         check_achievements(&stats_copy, &mut self.achievements)
     }
@@ -462,7 +489,14 @@ fn is_consecutive_day(prev_date: &str, current_date: &str) -> bool {
 
 // Pricing as of February 2026
 // https://platform.claude.com/docs/en/about-claude/models/overview
-pub fn calculate_cost(input_tokens: u64, output_tokens: u64, model: &str) -> f64 {
+// Cache pricing: https://platform.claude.com/docs/en/build-with-claude/prompt-caching
+pub fn calculate_cost(
+    input_tokens: u64,
+    output_tokens: u64,
+    model: &str,
+    cache_creation_tokens: Option<u64>,
+    cache_read_tokens: Option<u64>,
+) -> f64 {
     let (input_price_per_million, output_price_per_million) = match model {
         // Current generation (Claude 4.5)
         "claude-opus-4-5-20251101" => (5.0, 25.0),
@@ -487,10 +521,25 @@ pub fn calculate_cost(input_tokens: u64, output_tokens: u64, model: &str) -> f64
         _ => (3.0, 15.0),
     };
 
+    // Regular input/output tokens
     let input_cost = (input_tokens as f64 / 1_000_000.0) * input_price_per_million;
     let output_cost = (output_tokens as f64 / 1_000_000.0) * output_price_per_million;
 
-    input_cost + output_cost
+    // Cache write tokens (cache creation) cost 1.25x the base input price
+    let cache_write_cost = if let Some(cache_creation) = cache_creation_tokens {
+        (cache_creation as f64 / 1_000_000.0) * input_price_per_million * 1.25
+    } else {
+        0.0
+    };
+
+    // Cache read tokens cost 0.1x (10%) the base input price
+    let cache_read_cost = if let Some(cache_read) = cache_read_tokens {
+        (cache_read as f64 / 1_000_000.0) * input_price_per_million * 0.1
+    } else {
+        0.0
+    };
+
+    input_cost + output_cost + cache_write_cost + cache_read_cost
 }
 
 #[derive(Debug, Clone, Serialize, Deserialize)]
@@ -609,7 +658,7 @@ mod tests {
 
     #[test]
     fn test_cost_calculation_sonnet() {
-        let cost = calculate_cost(1000, 2000, "claude-sonnet-4-20250514");
+        let cost = calculate_cost(1000, 2000, "claude-sonnet-4-20250514", None, None);
         // 1000 input * $3/M = $0.003
         // 2000 output * $15/M = $0.030
         // Total = $0.033
@@ -618,7 +667,7 @@ mod tests {
 
     #[test]
     fn test_cost_calculation_opus() {
-        let cost = calculate_cost(1000, 2000, "claude-opus-4-20250514");
+        let cost = calculate_cost(1000, 2000, "claude-opus-4-20250514", None, None);
         // 1000 input * $15/M = $0.015
         // 2000 output * $75/M = $0.150
         // Total = $0.165
@@ -627,7 +676,7 @@ mod tests {
 
     #[test]
     fn test_cost_calculation_opus_45() {
-        let cost = calculate_cost(1000, 2000, "claude-opus-4-5-20251101");
+        let cost = calculate_cost(1000, 2000, "claude-opus-4-5-20251101", None, None);
         // Opus 4.5 pricing: $5/MTok input, $25/MTok output
         // 1000 input tokens = $0.005, 2000 output tokens = $0.05
         // Total = $0.055
@@ -636,7 +685,7 @@ mod tests {
 
     #[test]
     fn test_cost_calculation_haiku() {
-        let cost = calculate_cost(1000, 2000, "claude-3-5-haiku-20241022");
+        let cost = calculate_cost(1000, 2000, "claude-3-5-haiku-20241022", None, None);
         // 1000 input * $1/M = $0.001
         // 2000 output * $5/M = $0.010
         // Total = $0.011
@@ -645,14 +694,14 @@ mod tests {
 
     #[test]
     fn test_cost_calculation_unknown_defaults_to_sonnet() {
-        let cost = calculate_cost(1000, 2000, "some-unknown-model");
+        let cost = calculate_cost(1000, 2000, "some-unknown-model", None, None);
         // Should default to Sonnet pricing
         assert!((cost - 0.033).abs() < 0.0001);
     }
 
     #[test]
     fn test_cost_calculation_legacy_sonnet() {
-        let cost = calculate_cost(1000, 2000, "claude-3-5-sonnet-20241022");
+        let cost = calculate_cost(1000, 2000, "claude-3-5-sonnet-20241022", None, None);
         // Same as Sonnet 4 pricing
         assert!((cost - 0.033).abs() < 0.0001);
     }
@@ -660,7 +709,7 @@ mod tests {
     #[test]
     fn test_usage_stats_accumulation() {
         let mut stats = UsageStats::new();
-        stats.add_usage(1000, 2000, "claude-sonnet-4-20250514");
+        stats.add_usage(1000, 2000, "claude-sonnet-4-20250514", None, None);
 
         assert_eq!(stats.total_input_tokens, 1000);
         assert_eq!(stats.total_output_tokens, 2000);
@@ -672,8 +721,8 @@ mod tests {
     #[test]
     fn test_usage_stats_multiple_accumulations() {
         let mut stats = UsageStats::new();
-        stats.add_usage(1000, 1000, "claude-sonnet-4-20250514");
-        stats.add_usage(500, 500, "claude-sonnet-4-20250514");
+        stats.add_usage(1000, 1000, "claude-sonnet-4-20250514", None, None);
+        stats.add_usage(500, 500, "claude-sonnet-4-20250514", None, None);
 
         assert_eq!(stats.total_input_tokens, 1500);
         assert_eq!(stats.total_output_tokens, 1500);
@@ -684,17 +733,17 @@ mod tests {
     #[test]
     fn test_usage_stats_model_updated() {
         let mut stats = UsageStats::new();
-        stats.add_usage(1000, 1000, "claude-sonnet-4-20250514");
+        stats.add_usage(1000, 1000, "claude-sonnet-4-20250514", None, None);
         assert_eq!(stats.model, Some("claude-sonnet-4-20250514".to_string()));
 
-        stats.add_usage(500, 500, "claude-opus-4-20250514");
+        stats.add_usage(500, 500, "claude-opus-4-20250514", None, None);
         assert_eq!(stats.model, Some("claude-opus-4-20250514".to_string()));
     }
 
     #[test]
     fn test_session_reset() {
         let mut stats = UsageStats::new();
-        stats.add_usage(1000, 2000, "claude-sonnet-4-20250514");
+        stats.add_usage(1000, 2000, "claude-sonnet-4-20250514", None, None);
         stats.reset_session();
 
         assert_eq!(stats.total_input_tokens, 1000);
@@ -921,7 +970,7 @@ mod tests {
     #[test]
     fn test_usage_stats_serialization() {
         let mut stats = UsageStats::new();
-        stats.add_usage(1000, 2000, "claude-sonnet-4-20250514");
+        stats.add_usage(1000, 2000, "claude-sonnet-4-20250514", None, None);
         stats.increment_messages();
 
         // UsageStats should be serializable (for events)
@@ -950,7 +999,7 @@ mod tests {
     #[test]
     fn test_stats_update_event_serialization() {
         let mut stats = UsageStats::new();
-        stats.add_usage(100, 200, "claude-sonnet-4-20250514");
+        stats.add_usage(100, 200, "claude-sonnet-4-20250514", None, None);
 
         let event = StatsUpdateEvent { stats };
         let json = serde_json::to_string(&event).expect("Failed to serialize");
@@ -1004,7 +1053,7 @@ mod tests {
     #[test]
     fn test_context_tracking_update() {
         let mut stats = UsageStats::new();
-        stats.add_usage(50_000, 10_000, "claude-sonnet-4-20250514");
+        stats.add_usage(50_000, 10_000, "claude-sonnet-4-20250514", None, None);
 
         assert_eq!(stats.context_tokens_used, 50_000);
         assert_eq!(stats.context_window_limit, 200_000);
@@ -1014,8 +1063,8 @@ mod tests {
     #[test]
     fn test_context_tracking_accumulates() {
         let mut stats = UsageStats::new();
-        stats.add_usage(50_000, 10_000, "claude-sonnet-4-20250514");
-        stats.add_usage(50_000, 10_000, "claude-sonnet-4-20250514");
+        stats.add_usage(50_000, 10_000, "claude-sonnet-4-20250514", None, None);
+        stats.add_usage(50_000, 10_000, "claude-sonnet-4-20250514", None, None);
 
         assert_eq!(stats.context_tokens_used, 100_000);
         assert!((stats.context_utilisation_percent - 50.0).abs() < 0.1);
@@ -1079,7 +1128,7 @@ mod tests {
     #[test]
     fn test_context_reset_on_session_reset() {
         let mut stats = UsageStats::new();
-        stats.add_usage(100_000, 20_000, "claude-sonnet-4-20250514");
+        stats.add_usage(100_000, 20_000, "claude-sonnet-4-20250514", None, None);
 
         assert!(stats.context_tokens_used > 0);
         assert!(stats.context_utilisation_percent > 0.0);

src-tauri/src/types.rs

@@ -4,6 +4,10 @@ use serde::{Deserialize, Serialize};
 pub struct UsageInfo {
     pub input_tokens: u64,
     pub output_tokens: u64,
+    #[serde(default)]
+    pub cache_creation_input_tokens: Option<u64>,
+    #[serde(default)]
+    pub cache_read_input_tokens: Option<u64>,
 }
 
 #[derive(Debug, Clone, Serialize, Deserialize, PartialEq, Default)]

src-tauri/src/wsl_bridge.rs

@@ -355,6 +355,15 @@ impl WslBridge {
     pub fn send_message(&mut self, message: &str) -> Result<(), String> {
         let stdin = self.stdin.as_mut().ok_or("Process not running")?;
 
+        // Track input for cost estimation on interrupt
+        {
+            let mut stats = self.stats.write();
+            stats.current_request_input = Some(message.to_string());
+            stats.current_request_output_chars = 0;
+            stats.current_request_thinking_chars = 0;
+            stats.current_request_tools.clear();
+        }
+
         let input = serde_json::json!({
             "type": "user",
             "message": {
@@ -419,6 +428,9 @@ impl WslBridge {
         // we have to kill the process. This is the only reliable way to stop it.
         // See: https://github.com/anthropics/claude-code/issues/3455
         if let Some(mut process) = self.process.take() {
+            // Estimate cost for interrupted request before killing
+            self.estimate_interrupted_request_cost(app);
+
             // Kill the process immediately
             let _ = process.kill();
             let _ = process.wait();
@@ -426,6 +438,15 @@ impl WslBridge {
             // Clear stdin
             self.stdin = None;
 
+            // Clear tracking fields
+            {
+                let mut stats = self.stats.write();
+                stats.current_request_input = None;
+                stats.current_request_output_chars = 0;
+                stats.current_request_thinking_chars = 0;
+                stats.current_request_tools.clear();
+            }
+
             // Keep session_id and working directory for user reference
             // The user will see what session was interrupted
 
@@ -442,6 +463,99 @@ impl WslBridge {
         }
     }
 
+    fn estimate_interrupted_request_cost(&mut self, app: &AppHandle) {
+        // Read tracking data from stats
+        let (input_chars, output_chars, thinking_chars, tools, model) = {
+            let stats = self.stats.read();
+
+            // Only estimate if we have tracked content
+            if stats.current_request_input.is_none()
+                && stats.current_request_output_chars == 0
+                && stats.current_request_thinking_chars == 0
+                && stats.current_request_tools.is_empty() {
+                return;
+            }
+
+            let input_chars = stats.current_request_input.as_ref().map(|s| s.len() as u64).unwrap_or(0);
+            let model = stats.model.clone().unwrap_or_else(|| "claude-sonnet-4-5-20250929".to_string());
+
+            (input_chars, stats.current_request_output_chars, stats.current_request_thinking_chars, stats.current_request_tools.clone(), model)
+        };
+
+        println!("[COST ESTIMATION] Estimating cost for interrupted request");
+
+        // Use conservative 3.5 chars/token for estimation (vs standard 4)
+        let estimated_input_tokens = (input_chars as f64 / 3.5).ceil() as u64;
+        let estimated_output_tokens = ((output_chars as f64 / 3.5).ceil() as u64)
+            + ((thinking_chars as f64 / 3.5).ceil() as u64);
+
+        // Add tool overhead based on session averages
+        let mut tool_overhead_tokens = 0u64;
+        {
+            let stats = self.stats.read();
+            for tool_name in &tools {
+                if let Some(tool_stats) = stats.session_tools_usage.get(tool_name) {
+                    if tool_stats.call_count > 0 {
+                        // Use session average tokens per call for this tool
+                        let avg_tokens = (tool_stats.estimated_input_tokens + tool_stats.estimated_output_tokens)
+                            / tool_stats.call_count;
+                        tool_overhead_tokens += avg_tokens;
+                        println!("[COST ESTIMATION] Tool {} average: {} tokens", tool_name, avg_tokens);
+                    }
+                }
+            }
+        }
+
+        let total_estimated_input = estimated_input_tokens + tool_overhead_tokens;
+        let total_estimated_output = estimated_output_tokens;
+
+        // Add 20% safety margin to overestimate
+        let safety_margin = 1.2;
+        let conservative_input = (total_estimated_input as f64 * safety_margin).ceil() as u64;
+        let conservative_output = (total_estimated_output as f64 * safety_margin).ceil() as u64;
+
+        println!("[COST ESTIMATION] Input: {} chars → {} tokens (+ {} tool overhead) × 1.2 safety = {} tokens",
+            input_chars, estimated_input_tokens, tool_overhead_tokens, conservative_input);
+        println!("[COST ESTIMATION] Output: {} chars → {} tokens × 1.2 safety = {} tokens",
+            output_chars + thinking_chars,
+            estimated_output_tokens, conservative_output);
+
+        // Calculate cost (no cache tokens for interrupted requests)
+        let estimated_cost = calculate_cost(
+            conservative_input,
+            conservative_output,
+            &model,
+            None,
+            None,
+        );
+
+        println!("[COST ESTIMATION] Estimated cost: ${:.4} (conservative)", estimated_cost);
+
+        // Add to stats with estimated flag
+        {
+            let mut stats_guard = self.stats.write();
+            stats_guard.add_usage(
+                conservative_input,
+                conservative_output,
+                &model,
+                None,
+                None,
+            );
+        }
+
+        // Emit stats update
+        let stats_update_event = StatsUpdateEvent {
+            stats: self.stats.read().clone(),
+        };
+        let _ = app.emit("claude:stats", stats_update_event);
+
+        // Record to historical cost tracking (mark as estimated)
+        let app_clone = app.clone();
+        tauri::async_runtime::spawn(async move {
+            record_cost(&app_clone, conservative_input, conservative_output, estimated_cost).await;
+        });
+    }
+
     pub fn stop(&mut self, app: &AppHandle) {
         if let Some(mut process) = self.process.take() {
             let _ = process.kill();
@@ -603,48 +717,72 @@ fn process_json_line(
 
             // Only update stats if we have usage information
             if let Some(usage) = &message.usage {
-                if let Some(model) = &message.model {
-                    // Calculate cost for historical tracking
-                    let cost_usd = calculate_cost(usage.input_tokens, usage.output_tokens, model);
+                // Get model from message, or fall back to last known model from stats
+                let model = message.model.clone().or_else(|| {
+                    let stats_guard = stats.read();
+                    stats_guard.model.clone()
+                }).unwrap_or_else(|| {
+                    println!("[WARNING] No model info available for cost calculation, using default");
+                    "claude-sonnet-4-5-20250929".to_string()
+                });
 
-                    // Store cost for later use in output events
-                    message_cost = Some(MessageCost {
-                        input_tokens: usage.input_tokens,
-                        output_tokens: usage.output_tokens,
-                        cost_usd,
-                    });
+                // Calculate cost for historical tracking (including cache tokens)
+                let cost_usd = calculate_cost(
+                    usage.input_tokens,
+                    usage.output_tokens,
+                    &model,
+                    usage.cache_creation_input_tokens,
+                    usage.cache_read_input_tokens,
+                );
 
-                    // Batch all stats updates in a single write lock
-                    {
-                        let mut stats_guard = stats.write();
-                        stats_guard.increment_messages();
-                        stats_guard.add_usage(usage.input_tokens, usage.output_tokens, model);
-                        stats_guard.get_session_duration();
+                println!("Assistant message tokens - input: {}, output: {}, cache_creation: {:?}, cache_read: {:?}, cost: ${:.4}",
+                    usage.input_tokens,
+                    usage.output_tokens,
+                    usage.cache_creation_input_tokens,
+                    usage.cache_read_input_tokens,
+                    cost_usd
+                );
 
-                        // Attribute tokens to tools if any tools were used in this message
-                        if !tools_in_message.is_empty() {
-                            let per_tool_input = usage.input_tokens / tools_in_message.len() as u64;
-                            let per_tool_output = usage.output_tokens / tools_in_message.len() as u64;
-                            for tool in &tools_in_message {
-                                stats_guard.add_tool_tokens(tool, per_tool_input, per_tool_output);
-                            }
+                // Store cost for later use in output events
+                message_cost = Some(MessageCost {
+                    input_tokens: usage.input_tokens,
+                    output_tokens: usage.output_tokens,
+                    cost_usd,
+                });
+
+                // Batch all stats updates in a single write lock
+                {
+                    let mut stats_guard = stats.write();
+                    stats_guard.increment_messages();
+                    stats_guard.add_usage(
+                        usage.input_tokens,
+                        usage.output_tokens,
+                        &model,
+                        usage.cache_creation_input_tokens,
+                        usage.cache_read_input_tokens,
+                    );
+                    stats_guard.get_session_duration();
+
+                    // Attribute tokens to tools if any tools were used in this message
+                    if !tools_in_message.is_empty() {
+                        let per_tool_input = usage.input_tokens / tools_in_message.len() as u64;
+                        let per_tool_output = usage.output_tokens / tools_in_message.len() as u64;
+                        for tool in &tools_in_message {
+                            stats_guard.add_tool_tokens(tool, per_tool_input, per_tool_output);
                         }
                     }
-
-                    // Record to historical cost tracking
-                    let app_clone = app.clone();
-                    let input = usage.input_tokens;
-                    let output = usage.output_tokens;
-                    tauri::async_runtime::spawn(async move {
-                        record_cost(&app_clone, input, output, cost_usd).await;
-                    });
-
-                    // Don't emit here - we'll emit on Result message instead
-                    // This reduces the frequency of updates
-                } else {
-                    // Just increment message count if no usage info
-                    stats.write().increment_messages();
                 }
+
+                // Record to historical cost tracking
+                let app_clone = app.clone();
+                let input = usage.input_tokens;
+                let output = usage.output_tokens;
+                tauri::async_runtime::spawn(async move {
+                    record_cost(&app_clone, input, output, cost_usd).await;
+                });
+
+                // Don't emit here - we'll emit on Result message instead
+                // This reduces the frequency of updates
             } else {
                 // Just increment message count if no usage info
                 stats.write().increment_messages();
@@ -722,6 +860,14 @@ fn process_json_line(
         ClaudeMessage::StreamEvent { event } => {
             if event.event_type == "content_block_start" {
                 if let Some(block) = &event.content_block {
+                    // Track tool calls for cost estimation
+                    if block.block_type == "tool_use" {
+                        if let Some(name) = &block.name {
+                            let mut stats_guard = stats.write();
+                            stats_guard.current_request_tools.push(name.clone());
+                        }
+                    }
+
                     let state = match block.block_type.as_str() {
                         "thinking" => CharacterState::Thinking,
                         "text" => CharacterState::Typing,
@@ -739,7 +885,16 @@ fn process_json_line(
             } else if event.event_type == "content_block_delta" {
                 if let Some(delta) = &event.delta {
                     if let Some(text) = &delta.text {
+                        // Track output characters for cost estimation
+                        {
+                            let mut stats_guard = stats.write();
+                            stats_guard.current_request_output_chars += text.len() as u64;
+                        }
                         let _ = app.emit("claude:stream", text.clone());
+                    } else if let Some(thinking) = &delta.thinking {
+                        // Track thinking characters for cost estimation
+                        let mut stats_guard = stats.write();
+                        stats_guard.current_request_thinking_chars += thinking.len() as u64;
                     }
                 }
             }
@@ -768,12 +923,29 @@ fn process_json_line(
                     stats_guard.model.clone().unwrap_or_else(|| "claude-opus-4-20250514".to_string())
                 };
 
-                // Calculate cost for historical tracking
-                let cost_usd = calculate_cost(usage_info.input_tokens, usage_info.output_tokens, &model);
+                // Calculate cost for historical tracking (including cache tokens)
+                let cost_usd = calculate_cost(
+                    usage_info.input_tokens,
+                    usage_info.output_tokens,
+                    &model,
+                    usage_info.cache_creation_input_tokens,
+                    usage_info.cache_read_input_tokens,
+                );
 
                 let mut stats_guard = stats.write();
-                stats_guard.add_usage(usage_info.input_tokens, usage_info.output_tokens, &model);
-                println!("Result message tokens - input: {}, output: {}", usage_info.input_tokens, usage_info.output_tokens);
+                stats_guard.add_usage(
+                    usage_info.input_tokens,
+                    usage_info.output_tokens,
+                    &model,
+                    usage_info.cache_creation_input_tokens,
+                    usage_info.cache_read_input_tokens,
+                );
+                println!("Result message tokens - input: {}, output: {}, cache_creation: {:?}, cache_read: {:?}",
+                    usage_info.input_tokens,
+                    usage_info.output_tokens,
+                    usage_info.cache_creation_input_tokens,
+                    usage_info.cache_read_input_tokens
+                );
 
                 // Record to historical cost tracking
                 let app_clone = app.clone();
@@ -784,6 +956,15 @@ fn process_json_line(
                 });
             }
 
+            // Clear tracking fields since request completed successfully
+            {
+                let mut stats_guard = stats.write();
+                stats_guard.current_request_input = None;
+                stats_guard.current_request_output_chars = 0;
+                stats_guard.current_request_thinking_chars = 0;
+                stats_guard.current_request_tools.clear();
+            }
+
             // Always emit updated stats on result message (less frequent)
             // This includes the latest session duration
             let newly_unlocked = {

src/lib/commands/slashCommands.ts

@@ -37,6 +37,12 @@ async function changeDirectory(path: string): Promise<void> {
     // Capture conversation history before disconnecting
     const conversationHistory = claudeStore.getConversationHistory();
 
+    // Get currently granted tools and config auto-granted tools
+    const activeConversation = get(conversationsStore.activeConversation);
+    const grantedTools = activeConversation ? Array.from(activeConversation.grantedTools) : [];
+    const config = configStore.getConfig();
+    const allAllowedTools = [...new Set([...grantedTools, ...config.auto_granted_tools])];
+
     await invoke("stop_claude", { conversationId });
 
     // Wait for clean shutdown
@@ -50,12 +56,11 @@ async function changeDirectory(path: string): Promise<void> {
       conversationId,
       options: {
         working_dir: validatedPath,
+        allowed_tools: allAllowedTools,
       },
     });
 
     // Update Discord RPC when reconnecting after directory change
-    const config = configStore.getConfig();
-    const activeConversation = get(conversationsStore.activeConversation);
     if (activeConversation) {
       await updateDiscordRpc(
         activeConversation.name,
@@ -102,6 +107,12 @@ async function startNewConversation(): Promise<void> {
       conversationId,
     });
 
+    // Get granted tools before interrupting
+    const activeConversation = get(conversationsStore.activeConversation);
+    const grantedTools = activeConversation ? Array.from(activeConversation.grantedTools) : [];
+    const config = configStore.getConfig();
+    const allAllowedTools = [...new Set([...grantedTools, ...config.auto_granted_tools])];
+
     claudeStore.addLine("system", "Starting new conversation...");
     characterState.setState("thinking");
 
@@ -115,12 +126,11 @@ async function startNewConversation(): Promise<void> {
       conversationId,
       options: {
         working_dir: workingDir,
+        allowed_tools: allAllowedTools,
       },
     });
 
     // Update Discord RPC when starting new conversation
-    const config = configStore.getConfig();
-    const activeConversation = get(conversationsStore.activeConversation);
     if (activeConversation) {
       await updateDiscordRpc(
         activeConversation.name,

src/lib/components/CompactMode.svelte

@@ -5,6 +5,7 @@
   import { characterState, characterInfo } from "$lib/stores/character";
   import { isStreamerMode } from "$lib/stores/config";
   import { handleNewUserMessage } from "$lib/notifications/rules";
+  import { setSkipNextGreeting } from "$lib/tauri";
   import type { CharacterState, CharacterStateInfo } from "$lib/types/states";
 
   interface Props {
@@ -127,6 +128,9 @@
       const conversationId = get(claudeStore.activeConversationId);
       if (!conversationId) return;
 
+      // Set flag to preserve stats/permissions (don't treat next connect as new session)
+      setSkipNextGreeting(true);
+
       await invoke("interrupt_claude", { conversationId });
       claudeStore.addLine("system", "Interrupted");
       characterState.setState("idle");

src/lib/components/InputBar.svelte

@@ -338,23 +338,28 @@ User: ${formattedMessage}`;
             throw new Error("No active conversation");
           }
 
-          // Get current working directory before reconnecting
+          // Get current working directory and granted tools before reconnecting
           const workingDir = await invoke<string>("get_working_directory", { conversationId });
+          const activeConversation = get(conversationsStore.activeConversation);
+          const grantedTools = activeConversation
+            ? Array.from(activeConversation.grantedTools)
+            : [];
+          const config = configStore.getConfig();
+          const allAllowedTools = [...new Set([...grantedTools, ...config.auto_granted_tools])];
 
           // Set the flag to skip greeting on next connection
           setSkipNextGreeting(true);
 
-          // Reconnect to Claude
+          // Reconnect to Claude with preserved permissions
           await invoke("start_claude", {
             conversationId,
             options: {
               working_dir: workingDir,
+              allowed_tools: allAllowedTools,
             },
           });
 
           // Update Discord RPC when reconnecting
-          const config = configStore.getConfig();
-          const activeConversation = get(conversationsStore.activeConversation);
           if (activeConversation) {
             await updateDiscordRpc(
               activeConversation.name,

src/lib/components/PermissionModal.svelte

@@ -1,22 +1,17 @@
 <script lang="ts">
   import { invoke } from "@tauri-apps/api/core";
   import { get } from "svelte/store";
-  import { claudeStore, hasPermissionPending } from "$lib/stores/claude";
+  import { claudeStore } from "$lib/stores/claude";
   import { characterState } from "$lib/stores/character";
   import type { PermissionRequest } from "$lib/types/messages";
-  import { updateDiscordRpc } from "$lib/tauri";
+  import { updateDiscordRpc, setSkipNextGreeting } from "$lib/tauri";
   import { conversationsStore } from "$lib/stores/conversations";
   import { configStore } from "$lib/stores/config";
 
-  let isVisible = $state(false);
   let permission: PermissionRequest | null = $state(null);
   let grantedToolsList: string[] = $state([]);
   let workingDirectory = $state("");
 
-  hasPermissionPending.subscribe((pending) => {
-    isVisible = pending;
-  });
-
   claudeStore.pendingPermission.subscribe((perm) => {
     permission = perm;
     if (perm) {
@@ -54,6 +49,9 @@
           throw new Error("No active conversation");
         }
 
+        // Prevent stats reset on reconnection
+        setSkipNextGreeting(true);
+
         await invoke("stop_claude", { conversationId });
 
         // Small delay to ensure clean shutdown
@@ -125,7 +123,7 @@ Please continue where we left off and retry that action now that you have permis
   }
 
   function handleKeydown(event: KeyboardEvent) {
-    if (!isVisible || !permission) return;
+    if (!permission) return;
 
     if (event.key === "Enter") {
       event.preventDefault();
@@ -139,7 +137,7 @@ Please continue where we left off and retry that action now that you have permis
 
 <svelte:window onkeydown={handleKeydown} />
 
-{#if isVisible && permission}
+{#if permission}
   <div
     class="permission-overlay fixed inset-0 bg-black/70 flex items-center justify-center z-50 backdrop-blur-sm"
   >

src/lib/components/StatusBar.svelte

@@ -30,7 +30,7 @@
     createSummary,
     sanitizeForJson,
   } from "$lib/utils/conversationUtils";
-  import { updateDiscordRpc } from "$lib/tauri";
+  import { updateDiscordRpc, setSkipNextGreeting } from "$lib/tauri";
 
   const DISCORD_URL = "https://chat.nhcarrigan.com";
   const DONATE_URL = "https://donate.nhcarrigan.com";
@@ -190,6 +190,9 @@
         throw new Error("No active conversation");
       }
       await invoke("stop_claude", { conversationId });
+
+      // Clear granted permissions when user explicitly disconnects
+      claudeStore.revokeAllTools();
     } catch (error) {
       console.error("Failed to stop Claude:", error);
     }
@@ -248,6 +251,9 @@
           : sanitizedContent;
 
       // Step 1: Disconnect from Claude to reset context
+      // Prevent stats reset on reconnection
+      setSkipNextGreeting(true);
+
       if (connectionStatus === "connected") {
         await invoke("stop_claude", { conversationId: activeId });
       }

src/lib/components/UserQuestionModal.svelte

@@ -5,7 +5,7 @@
   import { claudeStore, hasQuestionPending } from "$lib/stores/claude";
   import { characterState } from "$lib/stores/character";
   import type { UserQuestionEvent } from "$lib/types/messages";
-  import { updateDiscordRpc } from "$lib/tauri";
+  import { updateDiscordRpc, setSkipNextGreeting } from "$lib/tauri";
   import { conversationsStore } from "$lib/stores/conversations";
   import { configStore } from "$lib/stores/config";
 
@@ -89,6 +89,9 @@
     claudeStore.clearQuestion();
 
     try {
+      // Prevent stats reset on reconnection
+      setSkipNextGreeting(true);
+
       await invoke("stop_claude", { conversationId });
 
       await new Promise((resolve) => setTimeout(resolve, 500));

src/routes/+page.svelte

@@ -8,6 +8,7 @@
     initializeDiscordRpc,
     stopDiscordRpc,
     updateDiscordRpc,
+    setSkipNextGreeting,
   } from "$lib/tauri";
   import { configStore, applyTheme, applyFontSize, isCompactMode } from "$lib/stores/config";
   import { initNotificationSync, cleanupNotificationSync } from "$lib/stores/notifications";
@@ -278,6 +279,9 @@
       const conversationId = get(claudeStore.activeConversationId);
       if (!conversationId) return;
 
+      // Set flag to preserve stats/permissions (don't treat next connect as new session)
+      setSkipNextGreeting(true);
+
       await invoke("interrupt_claude", { conversationId });
       claudeStore.addLine("system", "Process interrupted");
     } catch (error) {